Skill Issue: Karpathy on code agents (NoPriors, 2026-03)
Notes on the NoPriors interview where Karpathy says he hasn't typed a line of code since December, that every remaining failure feels like a skill issue rather than a capability ceiling, and that token throughput is the resource you should feel nervous about wasting. Plus the jaggedness caveat that undercuts the whole picture.
Source: Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI,
NoPriors, 20 March 2026. YouTube auto-transcript, archived under
sources/2026-07/karpathy-nopriors-skill-issue/.
This is the podcast the owner half-remembered when he wrote that operating agents badly is a skill issue. It is also, at four days before Anthropic's harness-design post, the same moment in time from outside a lab. Auto Research: taking the researcher out of the loop has its own page; this one is everything else.
The state change
Karpathy dates it precisely: in December something flipped, from writing 80% of his code by hand to writing perhaps 20%, and by March to not having typed a line since. His word for the resulting condition is psychosis, used affectionately and about a dozen times. The claim he makes about it is worth separating from the tone — that the default workflow of building software changed, that most people have not registered it, and that the change is recent enough that nobody knows what good looks like yet.
"Everything is skill issue"
The phrase carries a specific argument. When something doesn't work, his instinct is that the capability is there and he failed to string it together — a bad AGENTS.md, no decent memory tool, the wrong parallelisation. Not a model ceiling.
The corollary is a resource metric. He is nervous when a subscription goes unspent, because leftover quota means he did not maximise his token throughput; if Codex runs out he switches to Claude. He draws the analogy to being a PhD student with idle GPUs — the flops were the thing you felt guilty about wasting, and now it is tokens. His interviewer puts the general version well: engineering spent a decade not feeling compute-bound, and the binding constraint has moved back to the operator.
That is the framing worth taking, and it cuts both ways. "It's a skill issue" is an empowering diagnosis precisely because it is unfalsifiable — there is always another configuration you have not tried, which is also why he describes the whole thing as infinite and himself as anxious about it. Nothing in the interview distinguishes a real skill issue from a real ceiling, and he does not claim it does.
Macro actions
The concrete practice underneath the framing. The unit of manipulation is no longer a line or a function but a piece of functionality handed to one agent, with a second piece routed to another agent that won't collide with it, a third doing research, a fourth writing a plan. Review is proportional to how much you care about that particular code.
The image he points at is Peter Steinberg's monitor tiled with Codex sessions across ten checked-out repos, each task sized to about twenty minutes at high effort. Note the sizing — twenty minutes is short enough to keep several in flight by hand, which is a different regime from the multi-hour unattended runs Anthropic and scheduled runs are aiming at. Karpathy is describing a human operating many short loops, not a human absent from long ones. Both are in the interview and they are not the same workflow.
Claws: persistence as its own layer
His term for the layer above an agent session: something that keeps looping, has its own sandbox, acts on your behalf while you are not watching, and — the part he singles out — carries a real memory system rather than context compaction when the window fills. He is explicit that default agents do not have this and that open-source claw implementations do.
His own is a home-automation elf reachable over WhatsApp: it found the Sonos system by scanning the LAN, reverse-engineered the API from web searches, and now runs lights, HVAC, shades, pool and a camera that texts him when a delivery van pulls up. Six apps replaced by one natural-language surface. The interesting claim attached to it is not the demo but the inference — that a great deal of bespoke app UI exists only because there was no other way to reach the API, and agents dissolve the need for it.
He also thinks personality is load-bearing, which is the one design point here nobody else in this cluster raises. His reading: Claude's praise is calibrated well enough that he notices when a half-formed idea gets a flat response and a good one gets warmth, and he catches himself trying to earn it; Codex is dry and doesn't seem to care what is being built. Whether that is a feature or a manipulation is not examined.
Unpacked in The claw layer: an agent that persists when you close the lid, including where the demo's security story does not hold together and why "his runs on a box at home" is an inference the transcript does not support.
The jaggedness caveat
The counterweight, and he gives it real weight: the whole thing is "bursting at the seams", still rough, and pushing too far ahead makes it net not useful. He describes talking simultaneously to a brilliant lifelong systems programmer and a ten-year-old, and says he still gets badly frustrated by agents doing nonsensical things and by loops that go wrong and stay wrong.
His hypothesis for why is the most useful thing in the interview: the models are shaped by reinforcement learning, so they improve exactly where behaviour is verifiable — did the program run, did the test pass — and stall everywhere else. Nuance about what you meant, knowing when to ask a clarifying question, anything soft: outside the optimised region. You are either on rails and moving at speed, or off them and meandering.
His evidence is a joke. Ask a state-of-the-art model for one and you get the same atoms joke you got four years ago, unchanged, while the same model will move mountains on an agentic task for hours. Capability in verifiable domains is visibly not generalising to everything else. That is a direct, cheap counter-example to the "it all comes along for free as models scale" story, and it is the same boundary Auto Research: taking the researcher out of the loop runs into from the other direction: what can be graded gets better, and what can't, doesn't.
Jaggedness: what RL optimises, and what stalls follows that through the cluster — it turns out to be the mechanism behind several findings this notebook had collected separately, including why the model is a poor QA agent out of the box. Two notes on the joke: Karpathy dates it at three or four years, his interviewer at five, and the transcript has no speaker labels, so the attribution is worth checking before it is quoted anywhere else.
Everything else, briefly
The interview ranges wide. The parts that touch this notebook:
- Speciation. The labs currently pursue one monolithic model good at everything. He expects specialisation eventually — smaller models with the cognitive core intact, tuned for a niche — but says the science of "manipulating the brains" is undeveloped: context windows are cheap and safe to work with, weights are not, and fine-tuning without losing capability is still not a reliable primitive. Relevant to the fine-tuning thread, and taken further in Speciation, and why we only ever touch the context window together with the open-source point below.
- Open source runs six to eight months behind, and that is fine. His analogy is Linux: the industry has always wanted a common open platform it feels safe building on, and the gap is a decent power balance rather than a failure. He is explicit that his objection to a fully closed frontier is about centralisation having a poor historical track record, not about capability.
- Jevons paradox for software. Software was scarce because it was expensive; making it cheap should increase demand for it, on the ATMs-and-bank-tellers pattern. Cautiously optimistic, explicitly hard to forecast.
- Digital first, atoms later. Bits move a million times more easily than matter, so expect an enormous amount of rewiring in digital work before the physical world moves, and expect the interesting companies in between to be sensors and actuators — the interface that lets a system ask the universe a question.
- Documentation is for agents now. Instead of HTML docs for humans, markdown for agents, because an agent will re-explain any part of it to any human at their level with infinite patience. He tried to make a walkthrough video for microGPT (an LLM boiled down to ~200 lines) and abandoned it as not adding much. His framing of what is left for a person: the few bits an agent cannot come up with — he notes an agent cannot invent microGPT but understands it completely once shown — and a skill that scripts the curriculum through it. That is the same bet as this notebook and as "what the agent can't see doesn't exist", arrived at from teaching rather than from engineering. He makes the same move earlier about smart-home apps — the argument, and where it thins out, is in The customer is not the human anymore.
Where he is speaking from
Worth recording because it affects how to read the rest: he is outside the frontier labs by choice, and names the trade-off honestly in both directions. Inside, you cannot be a free agent — there are things you can't say and things the organisation would like you to say, and at the moment stakes get high an employee has little sway over what the entity does. Outside, your judgment drifts, because the systems are opaque and you no longer see what is coming. He expects to go back and forth. No part of this interview should be read as reporting on what any lab currently runs internally.
Linked from
- Agentic engineering: the work moves to the harnessThe emerging discipline around long-running coding agents — designing the scaffolding, feedback loops and environments that let an agent do reliable work, rather than writing the code yourself. Entry point for the harness-design cluster in this notebook.
- Auto Research: taking the researcher out of the loopKarpathy's name for an agent loop that improves a model without a human in it — an objective, a metric, boundaries, and go. What it did to a repo he had hand-tuned for two decades, the one constraint that decides where it applies, and the parallel version he has not solved.
- DoorDash on agentic commerce and DOT (NoPriors)Notes on the NoPriors interview with DoorDash's two co-founders — a natural-language ordering surface that half the time sends people to a merchant they have never used, a delivery robot built to a use case rather than to a capability, and an internal benchmark built to work out what a 20× rise in model spend actually bought.
- Harness design: prior work still to readThe pieces the 2026 harness-design posts build on — Anthropic's earlier long-running-agent harness, context engineering, building effective agents, the frontend design skill — captured as references because the article bodies could not be fetched from this session.
- Jaggedness: what RL optimises, and what stallsKarpathy's argument that models are advancing only where a reward can be computed, with a joke that hasn't changed in three or four years as the demonstration — and why that is the same boundary the harness cluster keeps running into from the other side.
- Loop engineering: from writing prompts to writing loopsThe claim that the unit of work is shifting from the prompt to the loop — Cherny, Osmani and Guzman all saying a version of it — and why the interesting question is not how to keep an agent running but who decides what "done" means.
- Speciation, and why we only ever touch the context windowKarpathy expects specialised models and does not see them arriving, and his explanation is that adjusting weights without losing capability is still an underdeveloped science while context windows just work. Plus where he puts the open-weight models — six to eight months behind, and better off there.
- The claw layer: an agent that persists when you close the lidKarpathy's name for the layer above a coding session — something that keeps looping in its own sandbox, remembers more than a compacted context, and answers on one messaging channel. The home-automation example he built, and why it is a partial answer to two open questions in this notebook.
- The customer is not the human anymoreKarpathy makes the same argument twice in one interview — smart-home apps should be APIs, docs should be markdown for agents rather than HTML for people — because in both cases an agent consumes the interface and routes to a human. Where that lands for this notebook, where it thins out, and what changes when a vendor does it deliberately.
- The harness is a skill issueA piece by Anton Leicht on sovereign AI sent me looking for how the frontier labs actually run their internal coding agents. The answer is that they are still experimenting — nobody has Auto Research — which means the harness is mine to build, and building it well is a skill worth training.
- The last hundred feet: building toward a use caseDoorDash's argument for why they had to build their own robot — the field builds a capability and then hunts for a problem — plus the edge cases nobody writes down at a desk, and the one incumbent data advantage in the interview that is actually load-bearing.
- The LLM-wiki pattern: the idea this notebook runs onAndrej Karpathy's pattern for an LLM-maintained personal wiki — kept verbatim at the repo root as LLM_Wiki.md — is the founding idea mantua.io itself instantiates. What the pattern claims, and where this repo's implementation matches or departs from it.